The research, led by Andrew Dumit and colleagues at Watershed Technology Inc. in San Francisco, addresses a problem that distinguishes LCI mapping from most standard machine learning benchmarks. In supervised classification, a misclassified item typically produces some anomalous signal that can be detected downstream. In LCI database mapping, by contrast, an incorrect mapping produces no such anomaly. A wrongly matched dataset will dutifully return an emissions number, and that number will look perfectly plausible in a report. The only reliable way to catch such errors has traditionally been item-level expert review, which is prohibitively expensive at the scale of modern corporate procurement data, where organizations may need to map hundreds of thousands of line items. The result has been an industry-wide reliance on category-average emission factors, which sacrifice the resolution that activity-based accounting is supposed to provide.

The team’s solution is a deliberately structured two-agent system that separates the act of proxy selection from the act of quality assessment through what the authors call an information barrier. The first agent, the mapper, proposes LCI database matches using iterative, tool-augmented retrieval, searching the database and refining its candidates much as a human practitioner would. The second agent, the judge, is architecturally prevented from seeing anything the mapper did. It observes only the original input item, the proposed activity, and the activity’s metadata, and then scores the proposed mapping along two independent dimensions: emissions similarity and material similarity. This enforced separation is not an implementation convenience but a methodological choice, designed to prevent the judge from inheriting the mapper’s biases or rationalizing its choices after the fact.

The performance gains reported in the study are striking. On an evaluation set of 1,039 items spanning seven product categories, the mapper achieved 90.7 percent defensible accuracy — defined as the share of items mapped to an option that a domain expert would approve — with zero abstentions. By comparison, retrieval-based baselines reached only 19 to 43 percent, and prior automated systems, while sometimes avoiding outright errors, abstained on 70 to 73 percent of items, effectively punting the hard decisions back to humans. Because the mapper always commits to an answer, every procurement line item receives a concrete lifecycle-based emission factor rather than a coarse category average, which is precisely what item-level carbon accounting requires.

Even more consequential is what the judge component accomplishes. Because proxy errors are silent, the practical value of any automated mapping system depends on how well it can tell its own confident successes from its quiet failures. At a simulated expert review budget of 20 percent, the judge captured 67 percent of all mapping errors, compared with only 37 to 40 percent for heuristic baselines. When the analysis focused on severe errors — cases where the chosen proxy’s emissions deviated by more than 100 percent from the correct value — the judge caught 74 percent of them. This means that organizations deploying the system can concentrate their scarce expert review capacity on exactly the items where a wrong proxy would most distort the reported footprint, rather than sampling randomly or reviewing everything.

The information barrier also yields a property that is rare in applied carbon accounting tools: calibration. Because the judge never sees the mapper’s reasoning, its quality scores function as an independent audit of each mapping, and these scores turn out to be well calibrated against actual correctness. In practical terms, the system can auto-accept 30 percent of its mappings while incurring an error rate of just 0.3 percent on that auto-accepted subset. This selectivity framework connects the work to a broader literature on selective classification and learning to defer to experts, in which models must know not only how to predict but when their predictions deserve trust. The authors note that large language models are often overconfident and biased self-evaluators when asked to grade their own outputs, which strengthens the case for a structurally independent assessor rather than self-reflection.

The technical architecture draws on several strands of recent research. The mapper’s iterative retrieval approach reflects agentic patterns such as ReAct, in which a language model interleaves reasoning steps with tool calls to ground its decisions in external data — in this case, the LCI database itself. The judge’s evaluation role builds on work on LLM-as-a-judge methods, while its scoring dimensions echo established lifecycle inventory practice, notably the data quality indicators introduced by Weidema and Wesnæs in the 1990s and subsequent proxy selection methodologies for choosing the most appropriate LCI dataset. Earlier machine learning applications in life cycle assessment largely focused on classification or regression tasks; the present work differs by treating the mapping problem as one of defensible proxy selection with explicit, auditable quality control, consistent with the requirements of ISO 14044.

The evaluation combined proprietary internal data from Watershed’s technical assessments with a public benchmark. A reproducible subset of 275 items from the Amazon Parakeet dataset of emission factor recommendations has been released, allowing outside researchers to compare their own systems on identical ground truth. The remainder of the evaluation data derives from proprietary procurement records and cannot be published, and the system code itself is proprietary to Watershed, whose carbon accounting products the company commercializes. All authors are Watershed employees, and the paper states that no external funding was received. These disclosures situate the work within a growing wave of industry-led research into AI-assisted sustainability measurement, alongside recent efforts to apply generative AI to product carbon footprint estimation and LCA data quality assessment.

The broader implications reach well beyond one company’s product pipeline. Accurate Scope 3 accounting has become a central battleground in corporate climate accountability, as regulators, investors, and voluntary disclosure frameworks increasingly demand granular, activity-based figures rather than spend-based approximations. By demonstrating that automated mapping can approach expert quality while providing calibrated signals for where human judgment is still needed, the study sketches a plausible division of labor between machines and domain experts at a scale that neither could achieve alone. If such systems mature, the long-standing trade-off between the resolution of a carbon footprint and the cost of producing it may finally begin to loosen, moving organizations from category averages toward genuinely item-level transparency across their supply chains.

The study arrives at a moment when the infrastructure for lifecycle inventory data itself is expanding. Databases such as ecoinvent, which releases documented change reports with each version update, and newer entrants like China’s HiQLCD, continue to grow in geographic and sectoral coverage, yet the fundamental mismatch between what companies purchase and what databases contain persists. This is why proxy selection has long been recognized as a methodological challenge in its own right, with earlier work proposing structured selection methodologies and expert elicitation to patch inventory gaps for specific product categories such as laundry detergents.

The reporting context also matters. Under the Greenhouse Gas Protocol’s Corporate Value Chain standard and its technical guidance, companies are expected to prioritize the Scope 3 categories most relevant to their sector, and sector-specific technical notes from disclosure platforms such as CDP reinforce that purchased goods and services rank highest in relevance for most industries. Spend-based estimation, which multiplies procurement spending by sector-average emission factors, satisfies reporting requirements cheaply but obscures the physical processes actually driving emissions, limiting the value of the resulting figures for procurement decisions and supplier engagement.

One caution raised in the machine learning literature concerns correlated errors: when multiple automated mappings fail in similar ways, headline accuracy figures can mask systematic biases within particular product categories or database regions. The authors’ emphasis on item-level expert annotation of defensible mappings, supported by a dedicated sustainability data advisory team, reflects an awareness that evaluation quality ultimately bounds the trustworthiness of any automated accounting pipeline built upon it.

Subject of Research: Quality-aware automated mapping of procurement items to lifecycle inventory databases for Scope 3 carbon accounting using a two-agent AI system

Article Title: Quality-aware automation for LCI database mapping

Article References: Dumit, A., Rao, K., Ulissi, S., Watson, S., Feintzeig, J., Joyce, P. J., & Bao, S. (2026). Quality-aware automation for LCI database mapping. Journal of Industrial Ecology. https://doi.org/10.1007/s44498-026-00157-2

Image Credits: AI Generated

DOI: 10.1007/s44498-026-00157-2

Keywords: life cycle assessment, Scope 3 emissions, LCI database mapping, carbon accounting, large language models, two-agent AI, proxy selection, quality assessment, selective classification, emission factors, supply chain, machine learning